NIF的内存管理接口为enif_alloc/enif_free。
erl_nif.c
void* enif_alloc(size_t size)
{
return erts_alloc_fnf(ERTS_ALC_T_NIF, (Uint) size);
}
erl_alloc.h
ERTS_ALC_INLINE
void *erts_alloc_fnf(ErtsAlcType_t type, Uint size)
{
return (*erts_allctrs[ERTS_ALC_T2A(type)].alloc)(
ERTS_ALC_T2N(type),
erts_allctrs[ERTS_ALC_T2A(type)].extra,
size);
}
可以看出NIF的内存分配将直接通过ERTS_ALC_T_NIF对应的虚拟机内存分配器ERTS_ALC_A_DRIVER分配内存,ERTS_ALC_A_DRIVER也是利用alloc_util框架实现的内存分配器,详细文档请阅读http://www.erlang.org/doc/man/erts_alloc.html。
void enif_free(void* ptr)
{
erts_free(ERTS_ALC_T_NIF, ptr);
}
void erts_free(ErtsAlcType_t type, void *ptr)
{
(*erts_allctrs[ERTS_ALC_T2A(type)].free)(
ERTS_ALC_T2N(type),
erts_allctrs[ERTS_ALC_T2A(type)].extra,
ptr);
}
对于NIF的内存释放过程也是如此,erlang虚拟机内存管理是一个非常庞杂的系统,此处将不进行分析,读者可以简单地将其看作malloc/free接口(虽然其实现要复杂的多)。
NIF的类型系统接口大同小异,基本上对于每种类型,都有一对make和get接口,稍微特殊的是binary类型。
首先来看NIF的利用进程堆分配内存的接口,它们是make类函数均要使用到的:
static ERTS_INLINE Eterm* alloc_heap(ErlNifEnv* env, unsigned need)
{
Eterm* hp = env->hp;
env->hp += need;
if (env->hp <= env->hp_end) {
return hp;
}
/* env的堆来自于其附着的进程的堆, 若env的堆有足够大的空间,则直接在堆内分配,否则将扩大堆 */
return alloc_heap_heavy(env, need, hp);
}
static Eterm* alloc_heap_heavy(ErlNifEnv* env, unsigned need, Eterm* hp)
{
env->hp = hp;
if (env->heap_frag == NULL) {
ASSERT(HEAP_LIMIT(env->proc) == env->hp_end);
HEAP_TOP(env->proc) = env->hp;
}
else {
env->heap_frag->used_size = hp - env->heap_frag->mem;
ASSERT(env->heap_frag->used_size <= env->heap_frag->alloc_size);
}
hp = erts_heap_alloc(env->proc, need, MIN_HEAP_FRAG_SZ);
/* 此处扩大进程的堆 */
env->heap_frag = MBUF(env->proc);
env->hp = hp + need;
env->hp_end = env->heap_frag->mem + env->heap_frag->alloc_size;
return hp;
}
Eterm*erts_heap_alloc(Process* p, Uint need, Uint xtra)
{
ErlHeapFragment* bp;
Eterm* htop;
Uint n;
n = need + xtra;
bp = MBUF(p);
if (bp != NULL && need <= (bp->alloc_size - bp->used_size)) {
Eterm* ret = bp->mem + bp->used_size;
bp->used_size += need;
return ret;
}
/* 进程的堆在开始时是和进程栈连在一起的,当堆不断扩大,直到不足时,分配器将为堆产生一个新的堆内存片段,之后的内存分配都将在新的堆内存片段上进行,这也是一种懒惰方法 */
bp = (ErlHeapFragment*)
ERTS_HEAP_ALLOC(ERTS_ALC_T_HEAP_FRAG, ERTS_HEAP_FRAG_SIZE(n));
/* 分配新的堆内存片段,使用ERTS_ALC_T_HEAP_FRAG对应的ERTS_ALC_A_EHEAP分配器分配内存,它也是一个通过alloc_util框架实现的内存分配器 */
htop = HEAP_TOP(p);
if (htop < HEAP_LIMIT(p)) {
*htop = make_pos_bignum_header(HEAP_LIMIT(p)-htop-1);
HEAP_TOP(p) = HEAP_LIMIT(p);
}
bp->next = MBUF(p);
MBUF(p) = bp;
/* 更新进程的堆内存片段信息,堆内存片段是一个单向列表,这也保证了进程堆的自由扩大 */
bp->alloc_size = n;
bp->used_size = need;
MBUF_SIZE(p) += n;
bp->off_heap.first = NULL;
bp->off_heap.overhead = 0;
return bp->mem;
}
#define ERTS_HEAP_ALLOC(Type, Size) \
erts_alloc((Type), (Size))
ERTS_ALC_INLINE void *erts_alloc(ErtsAlcType_t type, Uint size)
{
void *res;
res = (*erts_allctrs[ERTS_ALC_T2A(type)].alloc)(
ERTS_ALC_T2N(type),
erts_allctrs[ERTS_ALC_T2A(type)].extra,
size);
if (!res)
erts_alloc_n_enomem(ERTS_ALC_T2N(type), size);
return res;
}
对于一些常见的类型,其类型构建过程如下:
ERL_NIF_TERM enif_make_int(ErlNifEnv* env, int i)
{
#if SIZEOF_INT == ERTS_SIZEOF_ETERM
return IS_SSMALL(i) ? make_small(i) : small_to_big(i,
alloc_heap(env,2));
#elif (SIZEOF_LONG == ERTS_SIZEOF_ETERM) || \
(SIZEOF_LONG_LONG == ERTS_SIZEOF_ETERM)
return
make_small(i);
#endif
}
对于64位系统,无需为int分配内存,直接将数据内容放置在ERL_NIF_TERM中即可,对于32位大数字才需要分配内存,可见erlang虚拟机对内存分配已经到了抠门的地步了。
ERL_NIF_TERM enif_make_string(ErlNifEnv* env, const char* string, ErlNifCharEncoding encoding)
{
return
enif_make_string_len(env, string, sys_strlen(string), encoding);
}
ERL_NIF_TERM enif_make_string_len(ErlNifEnv* env, const char* string, size_t len, ErlNifCharEncoding encoding)
{
Eterm* hp =
alloc_heap(env,
len*2);
ASSERT(encoding == ERL_NIF_LATIN1);
return
erts_bld_string_n(&hp,NULL,string,len);
}
Eterm erts_bld_string_n(Uint **hpp, Uint *szp, const char *str, Sint len)
{
Eterm res = THE_NON_VALUE;
Sint i = len;
if (szp)
*szp += len*2;
if (hpp) {
res = NIL;
while (--i >= 0) {
res =
CONS(*hpp, make_small((byte) str[i]), res);
*hpp += 2;
}
}
return res;
}
string也是列表,因此需要分配两倍内存,一个用于保存指针,另一个用于保存数据,构建string时,需要逆序遍历原先的字符串数组。
ERL_NIF_TERM enif_make_tuple(ErlNifEnv* env, unsigned cnt, ...)
{
Eterm* hp =
alloc_heap(env,cnt+1);
Eterm ret = make_tuple(hp);
va_list ap;
*hp++ = make_arityval(cnt);
va_start(ap,cnt);
while (cnt--) {
*hp++ = va_arg(ap,Eterm);
}
va_end(ap);
return ret;
}
tuple是复合类型,仅仅需要在堆上分配tuple的元组个数+1个Eterm即可,一个用于保存tuple本身,其它的用于记录tuple每个成员。
ERL_NIF_TERM enif_make_list(ErlNifEnv* env, unsigned cnt, ...)
{
if (cnt == 0) {
return NIL;
}
else {
Eterm* hp =
alloc_heap(env,cnt*2);
Eterm ret = make_list(hp);
Eterm* last = &ret;
va_list ap;
va_start(ap,cnt);
while (cnt--) {
*last = make_list(hp);
*hp = va_arg(ap,Eterm);
last = ++hp;
++hp;
}
va_end(ap);
*last = NIL;
return ret;
}
}
list分配时也需要分配两倍内存,过程与string类似。
binary的构建有些特殊,分为两个阶段:分配与构造。
binary分配:
unsigned char* enif_make_new_binary(ErlNifEnv* env, size_t size,
ERL_NIF_TERM* termp)
{
flush_env(env);
*termp =
new_binary(env->proc, NULL, size);
/* 分配新的binary */
cache_env(env);
return binary_bytes(*termp);
}
Eterm new_binary(Process *p, byte *buf, Uint len)
{
ProcBin* pb;
Binary* bptr;
if (len <= ERL_ONHEAP_BIN_LIMIT) {
ErlHeapBin* hb = (ErlHeapBin *) HAlloc(p, heap_bin_size(len));
hb->thing_word = header_heap_bin(len);
hb->size = len;
if (buf != NULL) {
sys_memcpy(hb->data, buf, len);
}
return make_binary(hb);
}
/* 对于小于ERL_ONHEAP_BIN_LIMIT(64)字节的binary,可以直接分配在进程堆上 */
bptr =
erts_bin_nrml_alloc(len);
/* 对于大于ERL_ONHEAP_BIN_LIMIT(64)字节的binary,将通过ERTS_ALC_T_BINARY对应的ERTS_ALC_A_BINARY分配器进行分配, ERTS_ALC_A_BINARY也是利用alloc_util框架实现的内存分配器 */
bptr->flags = 0;
bptr->orig_size = len;
erts_refc_init(&bptr->refc, 1);
if (buf != NULL) {
sys_memcpy(bptr->orig_bytes, buf, len);
}
/* 然后构建一个进程binary的结构保存刚刚分配的大额binary */
pb = (ProcBin *) HAlloc(p, PROC_BIN_SIZE);
pb->thing_word = HEADER_PROC_BIN;
pb->size = len;
pb->next = MSO(p).first;
MSO(p).first = (struct erl_off_heap_header*)pb;
pb->val = bptr;
pb->bytes = (byte*) bptr->orig_bytes;
pb->flags = 0;
OH_OVERHEAD(&(MSO(p)), pb->size / sizeof(Eterm));
return make_binary(pb);
}
ERTS_GLB_INLINE Binary *erts_bin_nrml_alloc(Uint size)
{
Uint bsize = ERTS_SIZEOF_Binary(size) + CHICKEN_PAD;
void *res;
res = erts_alloc(ERTS_ALC_T_BINARY, bsize);
ERTS_CHK_BIN_ALIGNMENT(res);
return (Binary *) res;
}
binary构造:
Eterm enif_make_binary(ErlNifEnv* env, ErlNifBinary* bin)
{
if (bin->bin_term != THE_NON_VALUE) {
return bin->bin_term;
}
else if (bin->ref_bin != NULL) {
Binary* bptr = bin->ref_bin;
ProcBin* pb;
Eterm bin_term;
/* !! Copy-paste from new_binary() !! */
pb = (ProcBin *) alloc_heap(env, PROC_BIN_SIZE);
pb->thing_word = HEADER_PROC_BIN;
pb->size = bptr->orig_size;
pb->next = MSO(env->proc).first;
MSO(env->proc).first = (struct erl_off_heap_header*) pb;
pb->val = bptr;
pb->bytes = (byte*) bptr->orig_bytes;
pb->flags = 0;
OH_OVERHEAD(&(MSO(env->proc)), pb->size / sizeof(Eterm));
bin_term = make_binary(pb);
if (erts_refc_read(&bptr->refc, 1) == 1) {
/* Total ownership transfer */
bin->ref_bin = NULL;
bin->bin_term = bin_term;
}
return bin_term;
}
else {
flush_env(env);
bin->bin_term = new_binary(env->proc, bin->data, bin->size);
cache_env(env);
return bin->bin_term;
}
}
同样,对于get系列接口,也是大同小异的:
int enif_get_int(ErlNifEnv* env, Eterm term, int* ip)
{
#if SIZEOF_INT == ERTS_SIZEOF_ETERM
return term_to_Sint(term, (Sint*)ip);
#elif (SIZEOF_LONG == ERTS_SIZEOF_ETERM) || \
(SIZEOF_LONG_LONG == ERTS_SIZEOF_ETERM)
Sint i;
if (!term_to_Sint(term, &i) || i < INT_MIN || i > INT_MAX) {
return 0;
}
*ip = (int) i;
return 1;
#else
# error Unknown word size
#endif
}
对于64位系统和32位系统小数,直接可以从Eterm中提取数据内容,对于32位大数,需要一个较复杂的转换过程。
int enif_get_string(ErlNifEnv *env, ERL_NIF_TERM list, char* buf, unsigned len,
ErlNifCharEncoding encoding)
{
Eterm* listptr;
int n = 0;
ASSERT(encoding == ERL_NIF_LATIN1);
if (len < 1) {
return 0;
}
while (is_not_nil(list)) {
if (is_not_list(list)) {
buf[n] = '\0';
return 0;
}
listptr = list_val(list);
if (!is_byte(*listptr)) {
buf[n] = '\0';
return 0;
}
buf[n++] = unsigned_val(*listptr);
if (n >= len) {
buf[n-1] = '\0'; /* truncate */
return -len;
}
list = CDR(listptr);
}
buf[n] = '\0';
return n + 1;
}
取得string时,将重新拷贝一份。
int enif_get_tuple(ErlNifEnv* env, Eterm tpl, int* arity, const Eterm** array)
{
Eterm* ptr;
if (is_not_tuple(tpl)) {
return 0;
}
ptr = tuple_val(tpl);
*arity = arityval(*ptr);
*array = ptr+1;
return 1;
}
元组的取得较为简单,仅仅向调用者返回元组成员个数和元组成员数组。
int enif_get_list_cell(ErlNifEnv* env, Eterm term, Eterm* head, Eterm* tail)
{
Eterm* val;
if (is_not_list(term)) return 0;
val = list_val(term);
*head = CAR(val);
*tail = CDR(val);
return 1;
}
列表的取得过程比较麻烦,需要调用者遍历列表,不断对列表调用enif_get_list_cell直到最后一个元素。
int enif_inspect_binary(ErlNifEnv* env, Eterm bin_term, ErlNifBinary* bin)
{
ErtsAlcType_t allocator = is_proc_bound(env) ? ERTS_ALC_T_TMP : ERTS_ALC_T_NIF;
union {
struct enif_tmp_obj_t* tmp;
byte* raw_ptr;
}u;
u.tmp = NULL;
bin->data = erts_get_aligned_binary_bytes_extra(bin_term, &u.raw_ptr, allocator,
sizeof(struct enif_tmp_obj_t));
if (bin->data == NULL) {
return 0;
}
if (u.tmp != NULL) {
u.tmp->allocator = allocator;
u.tmp->next = env->tmp_obj_list;
u.tmp->dtor = &aligned_binary_dtor;
env->tmp_obj_list = u.tmp;
}
bin->bin_term = bin_term;
bin->size = binary_size(bin_term);
bin->ref_bin = NULL;
ADD_READONLY_CHECK(env, bin->data, bin->size);
return 1;
}
byte*erts_get_aligned_binary_bytes_extra(Eterm bin, byte** base_ptr, ErtsAlcType_t allocator, unsigned extra)
{
byte* bytes;
Eterm* real_bin;
Uint byte_size;
Uint offs = 0;
Uint bit_offs = 0;
if (is_not_binary(bin)) {
return NULL;
}
byte_size = binary_size(bin);
real_bin = binary_val(bin);
if (*real_bin == HEADER_SUB_BIN) {
ErlSubBin* sb = (ErlSubBin *) real_bin;
if (sb->bitsize) {
return NULL;
}
offs = sb->offs;
bit_offs = sb->bitoffs;
real_bin = binary_val(sb->orig);
}
if (*real_bin ==
HEADER_PROC_BIN) {
bytes = ((ProcBin *) real_bin)->bytes + offs;
} else {
bytes = (byte *)(&(((ErlHeapBin *) real_bin)->data)) + offs;
}
if (bit_offs) {
byte* buf = (byte *) erts_alloc(allocator, byte_size + extra);
*base_ptr = buf;
buf += extra;
erts_copy_bits(bytes, bit_offs, 1, buf, 0, 1, byte_size*8);
bytes = buf;
}
return bytes;
}
通常取得binary的时不会有数据拷贝,除非遇到通过匹配切分出的binary,这也是一种懒惰复制方法。
主要的类型系统接口已经分析完了,对于每个类型,都有一个get和make函数,binary类例外,get函数会取得类型的数据内容,make函数会为类型分配内存,并构造类型。